AI for Credit Unions (2026): NCUA Guidance, Governance, Use Cases and an AI Policy Outline

Where NCUA stands on AI in October 2026

NCUA’s position sits on one page, its artificial intelligence resource page, last modified on 28 April 2026. In its questions and answers the agency says:

• Credit unions may use AI tools and technologies, and NCUA supports adoption that is safe, sound and lawful.
• NCUA has not issued AI-specific rules or regulation. Existing rules are technology neutral, so information security standards apply whether a credit union communicates by email, by phone or through an AI tool.
• Examiners evaluate AI within the existing framework: safety and soundness, compliance with law, internal controls over the tool, ongoing risk monitoring, and due diligence on third parties. NCUA treats AI like any other new technology.
• When a credit union relies on an AI vendor, it must understand how the product works, the risks it introduces, how it fits the business model, and the vendor’s safeguards and controls. The board and management remain responsible for oversight.

The page links two older letters on third parties, 07-CU-13 on evaluating third-party relationships and 01-CU-20 on due diligence over service providers, plus outside resources from NIST, CISA, Treasury and FinCEN, including FinCEN’s alert on deepfake fraud schemes. It also invites credit unions to report regulatory barriers to AI through AskNCUA.

NCUA’s 2026 supervisory priorities letter, 26-CU-01, does not mention AI. Its priorities are balance sheet management and lending, payment systems and fraud prevention, and BSA/AML compliance, with third-party risk management reviewed where functions are outsourced. Most federal credit unions with assets of $50 million or less get defined-scope exams.

Model risk guidance is the open question. In May 2025 the Government Accountability Office (GAO-25-107197) found that the model risk material in NCUA’s examiner guide covers only interest rate risk modeling, and recommended that NCUA broaden it. NCUA generally agreed. At a July 2025 board briefing NCUA said model risk guidance alone would not give timely, use-case-specific direction, and that any new requirements would need rulemaking. GAO still lists the recommendation as open. The interagency model risk guidance banks follow, SR 26-2, is issued by the Federal Reserve, OCC and FDIC, not NCUA, though many credit unions borrow its structure.

NCUA’s own AI compliance plan and use case inventory

NCUA also publishes how it governs AI inside the agency. Its AI compliance plan, dated September 2025, responds to the AI in Government Act of 2020 and OMB Memorandum M-25-21. Each office submits its current and planned AI uses once a year for security, privacy and technical review; the Office of Business Innovation keeps the inventory; a Chief AI Officer reviews every use case; and four internal councils cover IT oversight, data governance, cybersecurity and enterprise risk. NCUA applies M-25-21’s minimum practices to high-impact AI and says systems that fail operational, ethical or regulatory requirements will be terminated. When the plan was written, the agency had deployed no public-facing AI.

The AI use case inventory is published as Excel and CSV files. Neither document binds credit unions, but the structure is a reasonable template for a credit union of any size: one inventory, an annual review, a named accountable officer and a process to retire tools that do not meet the bar.

The vendor and CUSO gap: NCUA cannot examine your AI provider

Most credit unions will buy AI rather than build it, often through their core processor, a fintech or a credit union service organization. That raises a structural point. NCUA’s authority to examine third-party vendors and CUSOs came from the Examination Parity and Year 2000 Readiness for Financial Institutions Act of 1998, was temporary, and expired at the end of 2001. NCUA can review vendors only with their permission, and GAO’s May 2025 report repeated its 2015 suggestion that Congress consider restoring the authority.

In September 2026 Representative Bill Foster introduced H.R. 10230, the Strengthening Oversight for the Financial Sector Act of 2026, which would give NCUA and the Federal Housing Finance Agency authority over the vendors their regulated entities use. The Defense Credit Union Council wrote to oppose it on 3 September 2026, arguing that it reaches well beyond the cybersecurity and AI concerns cited for it. When we checked on 22 October 2026, it had not been enacted.

For now the contract does the work an examiner cannot. For any AI vendor, require:

• a description of where member data is processed and stored, and whether prompts or documents are used to train models;
• security standards that meet Part 748 Appendix A, with evidence such as audit reports;
• breach notification fast enough for the credit union to meet its own 72-hour duty to NCUA;
• audit and access rights for the credit union, its auditors and, where the vendor agrees, NCUA examiners;
• data return and deletion on exit.

Part 748, GLBA and the 72-hour cyber incident rule

Three parts of 12 CFR Part 748 apply directly to AI:

• Section 748.0 requires a written security program that protects the security and confidentiality of member records, guards against unauthorized access that could cause substantial harm or serious inconvenience to a member, and responds to breaches.
• Appendix A, the guidelines for safeguarding member information that implement the Gramm-Leach-Bliley Act, expects board approval of the information security program, a risk assessment, controls such as access limits, encryption and monitoring, oversight of service providers by due diligence, contract and monitoring, and at least an annual report to the board. An AI tool that reads member data belongs in each of those steps.
• Appendix B sets out the response program for unauthorized access to member information, including member notice.

Section 748.1(c) adds the cyber incident notification rule: a federally insured credit union must notify NCUA as soon as possible and no later than 72 hours after it reasonably believes a reportable cyber incident has occurred. That includes incidents caused by a compromise at a CUSO, a cloud provider or another third-party data host. If a third party notifies the credit union first, the clock can start from that notice. NCUA’s 2026 cybersecurity report counted 588 cyber incident reports from credit unions between May 2025 and April 2026, and named AI-enabled phishing and agentic tools among the threats changing the landscape.

On assessment tools, the FFIEC retired its Cybersecurity Assessment Tool on 31 August 2025. NCUA continues to support its own Automated Cybersecurity Evaluation Toolbox (ACET), which maps to the FFIEC IT Examination Handbook and the NIST Cybersecurity Framework.

Fair lending and adverse action when AI touches credit

The Equal Credit Opportunity Act and Regulation B require a creditor that takes adverse action to give the applicant a statement of specific reasons. The CFPB’s Circular 2022-03 said this applies even when the decision comes from a complex algorithm, and Circular 2023-03 said a creditor cannot rely on the sample form’s checklist if those reasons do not reflect the factors actually used.

On 12 May 2025 the CFPB withdrew 67 guidance documents, and both circulars were on the list. The notice did not change Regulation B, and the CFPB said the withdrawal was not necessarily final. The duty to give accurate, specific reasons still comes from 12 CFR 1002.9. NCUA examines consumer compliance at credit unions with $10 billion in assets or less; the CFPB supervises larger ones.

In practice, keep generative AI out of the credit decision itself. Use it to assemble the file, extract income and collateral data with a link to the source page, and draft the officer’s memo. The approve-or-decline decision, and the reasons sent to the member, should come from a validated scoring process and a loan officer.

AI use cases for credit unions

Two of these carry the most risk. BSA narratives are confidential, so the system that drafts them must keep SAR information away from anyone without a need to know. Collections letters reach members in financial difficulty, so templates need compliance sign-off and the model should never invent payment terms.

Agentic AI for credit unions: where autonomy should stop

An agent differs from a chatbot because it acts: it calls the core system, opens a case, sends a message or updates a record. For a credit union the useful boundary is simple. Agents can read what the task needs and prepare work; a person approves anything that changes a member’s account, sends an external message, files a report or affects credit.

Three controls make that boundary real. Give each agent its own role with only the tools it needs. Put an approval step in front of every write action. Log every step, including the documents read and the tools called, so a supervisor or examiner can replay what happened. Agents also read untrusted content, such as uploaded documents and emails, which can carry hidden instructions, so an agent that reads member uploads should never hold permission to move money.

AI policy for credit unions: an outline the board can adopt

A board-approved AI policy turns NCUA’s general expectations into rules staff can follow. This outline covers what examiners are likely to ask about; adapt it to the credit union’s size. Our AI acceptable use policy template gives fuller wording for the staff-facing parts.

1. Purpose and scope. Which tools, staff, contractors and CUSOs the policy covers, and how it relates to the information security program under Part 748.
2. Roles. The board approves the policy; a named executive owns AI risk; IT, compliance, BSA and lending each own their use cases.
3. Approved tools and data classes. A table of approved tools against the member, employee and confidential data each may receive. Public consumer chatbots receive no member data.
4. Use case approval. Each use case gets an owner, a purpose, a risk tier, pre-launch testing on the credit union’s own documents and a review date. High-risk tiers include lending, collections, BSA and anything member-facing.
5. Human review. Which outputs a person must approve before use, and who that person is.
6. Fair lending and consumer protection. No generative AI in credit decisions without compliance review; adverse action reasons come from the decision process, not from a chatbot.
7. Vendors and CUSOs. Due diligence before contract, the contract terms listed above, and ongoing monitoring.
8. Security, logging and retention. Access control through the credit union’s identity system, prompt and output logs, and a retention period aligned with the records schedule.
9. Incidents. AI data exposure and AI-enabled fraud are handled under the incident response program, including the 72-hour NCUA notice.
10. Training and acceptable use. What staff may and may not paste into which tool, and how to spot deepfake and AI-generated fraud attempts.
11. Inventory and reporting. One AI inventory, reviewed at least annually, with a report to the board alongside the Appendix A security report.

Running governed AI on your own servers or a CUSO’s

Smaller credit unions often assume private AI needs a large IT team. In 2026 that is no longer true. Open-weight models in the 8 to 30 billion parameter range run on a single server-class GPU and handle summaries, policy questions and document extraction well, and a credit union can start with one workstation-class machine for a pilot.

Hosting the system on the credit union’s own infrastructure, or on infrastructure operated by a CUSO it already trusts with core processing, has three practical advantages under the rules above. Member data stays inside systems already covered by the Part 748 security program and its board reporting. The 72-hour incident clock depends on monitoring the credit union or its CUSO controls, rather than a vendor’s notice. And examiners can see the logs directly, which matters while NCUA cannot examine the vendor itself.

The trade-off is that someone has to run it. Many credit unions share that cost through a CUSO, the same way they share core processing and card services.

How VDF AI fits a credit union

VDF AI is software a credit union or its CUSO installs on infrastructure it controls, on-premises, in a private cloud or air-gapped. Models, retrieval indexes, conversation history and audit logs stay inside that environment, within the credit union’s Part 748 program. VDF AI holds no regulatory certification of its own; it is assessed as part of the credit union’s systems. A CUSO that serves several credit unions can run a separate installation for each.

For the use cases above, member service and staff policy assistants answer from the credit union’s own documents through permission-aware retrieval, so BSA files and HR records stay limited to the staff entitled to them. Governed agents prepare loan files and BSA case summaries and stop at a human approval step before anything is filed, sent or written to a member record. Every turn is logged with the user, model version, retrieved sources and tool calls. Staff sign in through Microsoft Entra ID single sign-on, which maps Entra groups to roles; other identity providers connect through an SSO-aware reverse proxy. Role-based access control sets which tools each role may use.

For more detail, see the finance and banking solution, the private GPT for banking guide, the banking AI use cases, the loan origination use case, the banking agents catalogue and the KYC and AML triage network playbook. Credit unions can begin with a three-month trial of the platform on their own servers.

Nothing here is legal advice; confirm specifics with your compliance counsel or NCUA examiner.